doi:10.7151/dmps.1141
ASYMPTOTIC BEHAVIOUR IN THE SET OF NONHOMOGENEOUS CHAINS OF STOCHASTIC
OPERATORS
1Ma lgorzata Pu lka Department of Mathematics Gda´ nsk University of Technology
ul. Gabriela Narutowicza 11/12, 80-233, Gda´ nsk, Poland e-mail: mpulka@mif.pg.gda.pl
Abstract
We study different types of asymptotic behaviour in the set of (infinite di- mensional) nonhomogeneous chains of stochastic operators acting on L
1(µ) spaces. In order to examine its structure we consider different norm and strong operator topologies. To describe the nature of the set of nonhomoge- neous chains of Markov operators with a particular limit behaviour we use the category theorem of Baire. We show that the geometric structure of the set of those stochastic operators which have asymptotically stationary density differs depending on the considered topologies.
Keywords: Markov operator, asymptotic stability, residuality, dense G
δ. 2010 Mathematics Subject Classification: Primary: 47A35, 47B65;
Secondary: 60J10, 54H20.
1. Introduction
The study of chains of Markov operators has become a subject of interest in regard to their applications in many different areas of science and technology. Markov operators are commonly used to describe phenomena involving a law of conserva- tion of a certain quantity, e.g. mass, energy, the number of particles in physical or chemical processes. Typical questions appear in the context of probability
1
This paper is a part of the author’s Ph.D. thesis written under the supervision of Professor
W. Bartoszek. The author wishes to express her appreciation to Professor Bartoszek for his
advice and helpful suggestions.
theory and concern the evolution of a density (probability distribution) of such a quantity. The case when the chain is homogeneous in time is well-understood and has a comprehensive literature (cf. [7, 8, 10]). In particular, the asymptotic behaviour of iterates of Markov operators has been intensively studied. The er- godic structure of homogeneous chains is fully described including probabilistic, lattice and spectral conditions for convergence of iterates with respect to all stan- dard topologies. In the case of the class of chains nonhomogeneous in time the situation is not so transparent, since no proper notion of a stationary density can be defined (in general). Thus, in order to describe the properties of a nonhomo- geneous chain one may study its asymptotic behaviour, which is understood as the study of a ”generalized concept” of stationarity. Namely, we may ask whether there exists a common limiting density or, at least, if the influence of the state of the process at the time m on its future states decreases to zero with the passage of time. Various gradations of this asymptotic properties may be considered de- pending on the mode of convergence of the iterates of the Markov operator. In this paper we focus solely on the uniform and strong modes of convergence.
Differences between the classes of homogeneous and nonhomogeneous chains attracted the attention of probabilists in the second half of the 20th century.
For example, in [6] Iosifescu observed that unlike the homogenous case, uniform asymptotic stability (strong ergodicity) is not a ”natural” concept for nonhomo- geneous chains. Thus, given a class of possible evolutions of Markov operators, i.e., a class of nonhomogeneous chains of Markov operators with a particular asymptotic behaviour, one may ask about its topological size. Such a description is based on the category theorem of Baire. Namely, the set is recognized as a large object if it is residual (it contains a dense G
δset). Thus, generic evolutions are those which belong to a residual subset. The aim of this paper is to define different types of asymptotic behaviour of nonhomogeneous chains of Markov op- erators acting on L
1(µ) spaces and to determine which one of them is prevalent.
The geometric structure of infinite dimensional nonhomogeneous Markov chains defined on the `
1space of all absolutely summable real sequences was intensively studied in [12]. Since `
1= L
1(N, 2
N, counting measure ), then the results included in this article are generalizations of the results obtained in [12].
For the convenience of the reader most of the theorems are proved in full detail.
This paper may be considered as the first step to generalizations of some results included in [3]. It is worth noticing that in [11] the asymptotic properties of nonhomogeneous discrete Markov processes with general state space L
1(µ) were studied and the results obtained were applied in the investigation of the limit be- haviour of the so-called quadratic stochastic processes which are concerned with genetic models.
Let (X, A, µ) be a separable σ-finite measure space. Throughout the paper
we consider the (separable) Banach lattice of real and A-measurable functions
f such that |f | is µ-integrable and we denote it by L
1(µ). By k · k
1we denote the relevant norm. We say that a linear operator P : L
1(µ) → L
1(µ) is Markov (or stochastic) if
P f ≥ 0 and kP f k
1= kf k
1for all f ≥ 0, f ∈ L
1(µ). By D = D(X, A, µ) we denote the set of all densities on X, i.e.,
D = f ∈ L
1(µ) : f ≥ 0, kf k
1= 1 .
In view of stochasticity of P we have that k|P |k = 1 (where k| · |k stands for the norm operator) and P (D) ⊂ D. The sequence of such operators denoted by P := (P
m,m+1)
m≥0is called a discrete time nonhomogeneous chain of Markov operators. For any natural numbers 0 ≤ m < n we set
P
m,n= P
m,m+1◦ P
m+1,m+2◦ · · · ◦ P
n−1,n.
If for each m ≥ 0 one has P
m,m+1= P , then P = (P )
m≥0is called a homogeneous chain of Markov operators. The set of all chains of Markov operators (including homogeneous) will be denoted by S , i.e.,
S = n
P = P
m,m+1m≥0
: P
m,m+1are Markov operators o .
Let t ∈ [0, 1] be given. A convex combination T(t) of two chains of Markov operators P and R ∈ S is defined as follows:
T
m,m+1(t) = tP
m,m+1+ (1 − t) R
m,m+1.
Note that T(t) ∈ S for every t ∈ [0, 1] and that a mapping [0, 1] 3 t 7→ T(t) ∈ S is continuous when S is equipped with suitable topology. Moreover, T(0) = R and T(1) = P. Thereby, S has an affine structure and it is arcwise connected.
Throughout the paper we write N
0= N ∪ {0}.
We will endow the set S with metric structures. Given P, R ∈ S let us consider:
(1) the sup norm operator topology induced by the metric d
n. sup(P, R) = sup
m
P
m,m+1− R
m,m+1, (2) the P norm operator topology induced by the metric
d
n.P(P, R) =
∞
X
m=0
1 2
m+1P
m,m+1− R
m,m+1,
(3) the P sup strong operator topology induced by the metric d
so. sup(P, R) =
∞
X
l=0
1 2
lsup
m
P
m,m+1f
l− R
m,m+1f
l1
,
where {f
0, f
1, . . .} is a fixed countable and linearly dense subset of D, (4) the P P strong operator topology induced by the metric
d
so.P(P, R) =
∞
X
m,l=0
1 2
m+l+1P
m,m+1f
l− R
m,m+1f
l1
,
where {f
0, f
1, . . .} is a fixed countable and linearly dense subset of D.
Note that d
so. sup(P
k, R) → 0 as k → ∞ if and only if for every f ∈ L
1(µ) (f ∈ D) and any m ∈ N
0one has lim
k→∞sup
mkP
km,m+1f − R
m,m+1f k
1= 0.
Moreover, the topologies generated by d
so. supand d
so.Pdo not depend on the choice of a sequence {f
0, f
1, . . .}.
Clearly, d
n. supgenerates the strongest topology and d
so.Pgenerates the weakest. However, it should be emphasized that metrics d
n.Pand d
so. supare not comparable. In order to observe it, consider P
j= (P
jm,m+1)
m≥0∈ S defined as follows:
P
jm,m+1=
( P, if 0 ≤ m < j, I, if m ≥ j,
where I stands for the identity operator and P = (P )
m≥0is such that P 6= I.
Then
d
n.P(P
j, P) =
∞
X
m=0
1 2
m+1P
jm,m+1− P
m,m+1=
j−1
X
m=0
1
2
m+1k|P − P |k +
∞
X
m=j
1
2
m+1k|I − P |k
= 1
2
jk|I − P |k → 0 as j → ∞.
On the other hand, for a given fixed countable and dense subset {f
0, f
1, . . .} of D we have
d
so. sup(P
j, P) =
∞
X
l=0
1 2
lsup
m
P
jm,m+1f
l− P
m,m+1f
l1
=
∞
X
l=0
1
2
lkf
l− P f
lk
1> 0.
Thus d
so. sup(P
j, P) 9 0 as j → ∞. Thereby, d
n.Pis not stronger than d
so. sup. Now we shall see that d
so. supis not stronger than d
n.P. In order to prove it observe that since the measure µ is σ-finite, there exists a sequence {B
k}, B
k∈ A, such that B
i∩ B
j= ∅ for i 6= j and
X =
∞
[
k=0
B
k, 0 < µ (B
k) < ∞ for all k ∈ N
0.
Let g
k∈ D be such that the essential support supp g
k:= {x ∈ X : g
k(x) 6= 0} ⊆ B
kfor any k ∈ N
0. For any f ∈ L
1(µ) denote
a
k(f ) = Z
Bk
f dµ.
Note that P
∞k=0
a
k(f ) = 1 if f ∈ D. Then we can define P
j= (P
jm,m+1)
j≥0∈ S as follows:
P
jm,m+1f = P
jf = f 1
Sjk=0Bk
+
∞
X
k=j+1
a
k(f ) · g
0.
Let I = (I, I, . . .) ∈ S , where as before I stands for the identity operator. Then
d
so. sup(P
j, I) =
∞
X
l=0
1 2
lsup
m
P
jm,m+1f
l− If
l1
=
∞
X
l=0
1
2
lkP
jf
l− f
lk
1=
∞
X
l=0
1 2
lf
l1
Sjk=0Bk
+
∞
X
k=j+1
a
k(f
l) g
0− f
l1
Sjk=0Bk
− f
l1
S∞k=j+1Bk
1
≤
∞
X
l=0
1 2
l
∞
X
k=j+1
a
k(f
l) kg
0k
1+ f
l1
S∞k=j+1Bk
1
=
∞
X
l=0
1 2
l
∞
X
k=j+1
a
k(f
l) + Z
S∞ k=j+1Bk
f
ldµ
=
∞
X
l=0
1 2
l· 2
∞
X
k=j+1
a
k(f
l)
=
∞
X
l=0
1
2
l−11 −
j
X
k=1
a
k(f
l)
!
→ 0 as j → ∞.
On the other hand, d
n.P(P
j, I) =
∞
X
m=0
1
2
m+1k|P
jm,m+1− I|k
=
∞
X
m=0
1 2
m+1!
k|P
j− I|k = 1 · 2 = 2 9 0 as j → ∞.
Therefore d
so. supis not stronger than d
n.P. It follows that the metrics d
n.Pand d
so. supare not comparable. The relationships between the considered metrics are illustrated in the Figure 1.
4
4 4 4
d
n. supd
so. supd
so.Pd
n.PFigure 1. The relationships between the metrics d
n. sup, d
n.P, d
so. sup, d
so.P.
Let us note that in the class of homogeneous chains of Markov operators, metrics d
n. supand d
n.Pare equivalent. In fact, if P = (P )
m≥0, R = (R)
m≥0∈ S , then d
n. sup(P, R) = d
n.P(P, R) = k|P − R|k. Similarly we find that in the homogeneous case metrics d
so. supand d
so.Pare equivalent and for any P = (P )
m≥0, R = (R)
m≥0∈ S one has d
so. sup(P, R) = d
so.P(P, R) = P
∞l=0 1
2l
kP f
l− Rf
lk
1, where {f
0, f
1, . . .} is a fixed countable and linearly dense subset of D. This supports our remark that the nonhomogeneous case is more complex than the homogeneous one.
In what follows we study different types of asymptotic behaviour of nonho-
mogeneous chains of stochastic operators as well as residualities in the set S .
We shall see that the geometric structure of the set of those stochastic operators
which have asymptotically stationary density differs depending on the considered
topologies. We prove that the set of those Markov operators which do not possess
limiting density is dense and its interior is nonempty in the topology induced by
the metric d
n. sup. On the other hand, it occurs that the set of those operators
for which the limiting density exists is dense while S is endowed with topology
induced by the metric d
n.P. We also examine the set of Markov operators which
we call (uniformly or strongly, if studied in norm or strong operator topology
respectively) almost asymptotically stable and we prove that it forms a residual
subset for both norm and strong operator topologies.
Note that d
n. supis the most relevant metric (topology) in studying the limit be- haviour of nonhomogeneous chains of stochastic operators. It should be empha- sized that, in contrast to the homogeneous case, the property of denseness of the set of nonhomogeneous chains of stochastic operators with a particular asymp- totic behaviour does not suffice to understand its ”size”. It derives from the fact that in the case of P sup and P P strong operator topologies the denseness of the complement of the set mentioned above can always be proved by modifying on the tail so-called sweeping operators or a fixed stochastic projection. There- fore, in order to describe the nature of the set we use the category theorem of Baire. This is because the space S equipped with any of the metrics (1)–(4) is complete and the classical Baire theorem is applicable.
The Baire category of asymptotic stability for homogeneous Markov chains was worked out in e.g. [2, 9, 13]. It should be clearly understood that our results are not a direct analogy of what was obtained in these works. In particular, it was proved in [13] that uniformly asymptotically stable (quasi-compact) ho- mogeneous Markov chains form a dense G
δsubset in norm operator topology.
In our nonhomogeneous case, the set of those chains of operators which are not uniformly asymptotically stable has a nonempty interior.
There are more relevant works in the literature dealing with the topic of the limit behaviour of nonhomogeneous chains of Markov operators. The reader should be warned that authors do not always use the same names for the same notions. For example, in [5] and [6] strong ergodicity is what we call uniform asymptotic stability and weak ergodicity is what we refer to as almost uniform asymptotic stability. Some authors apply the terminology derived from the er- godic theory to the theory of stochastic processes and denominate what we call asymptotic stability by mixing (cf. [2, 3, 9, 12]). See [5] for detailed classification of different types of asymptotic behaviour of nonhomogeneous Markov chains.
2. Uniform asymptotic stability
In this section we examine the strongest case of asymptotic stability of chains of Markov operators, i.e., uniform asymptotic stability. We start with
Definition. A nonhomogeneous chain of Markov operators P is called uniformly asymptotically stable if there exists a unique f
∗∈ D such that for every m ∈ N
0n→∞
lim sup
f ∈D
kP
m,nf − f
∗k
1= 0.
The set of all uniformly asymptotically stable chains of Markov operators is de-
noted by S
uas.
Note that uniformly asymptotically stable chains of operators possess common limiting density and the mode of convergence is uniform. The following theorem is concerned with the prevalence problem in the set S .
Theorem 1. The set S
uascof all Markov operators which are not uniformly asymptotically stable is a sup norm topology dense subset of S (i.e., in d
n. sup).
Moreover, in this case its interior IntS
uasc6= ∅.
Proof. Let P ∈ S and 0 < ε < 1 be taken arbitrarily. As before, since the measure µ is σ-finite, there exists a sequence {B
k}, B
k∈ A, such that B
i∩B
j= ∅ for i 6= j and
X =
∞
[
k=0
B
k, 0 < µ (B
k) < ∞ for all k ∈ N
0.
Let g
k∈ D be such that supp g
k= {x ∈ X : g
k(x) 6= 0} ⊆ B
kfor any k ∈ N
0. Then we can define R ∈ S as follows: for any f ∈ L
1(µ),
R
m,m+1f =
∞
X
j=m+1
a
j−m−1(f ) · g
j,
where a
j(f ) = R
Bj
f dµ. Note that P
∞j=0
a
j(f ) = 1 if f ∈ D. Consider a convex combination
P
εm,m+1= (1 − ε) P
m,m+1+ εR
m,m+1. Clearly, P
ε= (P
εm,m+1)
m≥0∈ S . We have
d
n.sup(P
ε, P) = sup
m
(1 − ε) P
m,m+1+ εR
m,m+1− P
m,m+1= ε sup
m
P
m,m+1− R
m,m+1≤ 2ε.
It remains to show that P
εis not uniformly asymptotically stable. Suppose that, on the contrary, there exists f
∗∈ D such that for every f ∈ D we have lim
n→∞P
εm,nf = f
∗. Since f ∈ D, there exists M ∈ N
0such that
Z
M
S
k=0
Bk
f
∗dµ > 1 − ε.
Hence
Z
M
S
k=0
Bk
P
εm,nf dµ −−−→
n→∞Z
M
S
k=0
Bk
f
∗dµ > 1 − ε.
On the other hand, if n > m > M , then Z
M
S
k=0
Bk
P
εm,n+1f dµ = 1 − Z
∞
S
k=M +1
Bk
P
εm,n+1f dµ
= 1 − Z
∞
S
k=M +1
Bk
(1 − ε) P
n,n+1+ εR
n,n+1(P
εm,nf ) dµ
≤ 1 − ε Z
∞
S
k=M +1
Bk
R
n,n+1(P
εm,nf ) dµ = 1 − ε.
It follows that S
uascis d
n. supdense in S .
To show that Int S
uasc6= ∅ for the sup norm topology (i.e., in d
n. sup) consider the open ball
K(R, 1) := {T ∈ S : d
n. sup(T, R) < 1} ,
where R is defined as before. If T ∈ K(R, 1), then for some 0 < ε < 1 sup
f ∈D
T
m,m+1f − R
m,m+1f
1
≤ d
n. sup(T, R) = 1 − ε.
Hence for every m + 1 > M and every f ∈ D, Z
M
S
k=0
Bk
T
0,m+1f dµ ≤ d
n. sup(T, R) = 1 − ε.
Thus,
sup
M ∈N
lim sup
m→∞
Z
M
S
k=0
Bk
T
0,m+1f dµ ≤ d
n. sup(T, R) = 1 − ε < 1,
and therefore T has no ”invariant” densities (common limiting density). It follows that T ∈ S
uasc.
In the next result we shall see that topologies on S generated by d
n. supand d
n.Pdiffer. Namely S
uasis large for d
n.P. In fact, we have
Proposition 2. The set S
uasis P norm topology dense in S (i.e., in d
n.P).
Proof. Let P ∈ S and 0 < ε < 1 be taken arbitrarily. There exists M ∈ N
0such that
21M< ε. Define P
ε∈ S as follows:
P
εm,m+1=
( P
m,m+1, if m ≤ M,
E, if m > M,
where Ef = ( R
X
f dµ)g for some fixed g ∈ D and any f ∈ L
1(µ). Obviously, E = (E
m,m+1)
m≥0∈ S , where for every m ∈ N
0, E
m,m+1= E. Then for every m ∈ N
0n→∞
lim k|P
εm,n− E|k = 0.
Therefore P
εis uniformly asymptotically stable. Clearly, d
n.P(P, P
ε) =
∞
X
m=M +1
1 2
m+1P
m,m+1− E ≤ 1
2
M< ε, which completes the proof.
We will now discuss a weaker case of asymptotic stability of chains of Markov operators, i.e., almost uniform asymptotic stability. We begin with
Definition. A nonhomogeneous chain of Markov operators P is said to be almost uniformly asymptotically stable if for every m ∈ N
0n→∞
lim sup
f,g∈D
kP
m,nf − P
m,ngk
1= 0.
The set of all almost uniformly asymptotically stable Markov operators is denoted by S
a.uas.
Repeating arguments from [1] or following the proof of Theorem 4.6 in [12] (cf.
[8]), we obtain a useful characterization of almost uniformly asymptotically stable nonhomogeneous chains of Markov operators.
Theorem 3. Let P ∈ S . If there exists a sequence (λ
n)
n∈N0, 0 ≤ λ
n< 1, satisfying
∞
X
n=0
λ
n= ∞
and such that for every f , g ∈ D we have P
n,n+1f ∧ P
n,n+1g
1
≥ λ
nfor all n ∈ N
0,
then P is almost uniformly asymptotically stable (here ∧ stands for the ordinary minimum in L
1(µ)).
Almost uniform asymptotic stability means that the influence of the state of the
process at the time m on its future states decreases (uniformly) to zero with the
passage of time. Thus, in the case of nonhomogeneous chains of Markov operators
this property is essentially weaker than the uniform asymptotic stability which
additionally claims the existence of a ”stationary” (common limiting) density. In the class of homogeneous chains of Markov operators notions of uniform asymp- totic stability and almost uniform asymptotic stability coincide. Indeed, if there exists ε > 0 such that for some n
0and every f , g ∈ D we have kP
n0f ∧P
n0gk
1≥ ε, then repeating arguments from [1] we obtain that kP
n0f −P
n0gk
1≤ (1−ε)kf −gk
1and we conclude that the mapping P is a strict contraction. Applying the Banach fixed point theorem there exists a unique P - invariant density f
∗∈ D such that lim
n→∞kP
nf − f
∗k
1= 0, where f ∈ D is arbitrary. It follows that
sup
f ∈D
kP
nf − f
∗k
1≤ (1 − ε)
jn n0
k
· kf − f
∗k ≤ 2 (1 − ε)
jn n0
k
→ 0
uniformly for f ∈ D. Hence in the class of homogeneous chains of Markov operators
S
a.uas= S
uas= {P ∈ S : ∃
n∃
ε>0∀
f,g∈DkP
nf ∧ P
ngk
1≥ ε} , which implies that in the homogeneous case the set S
a.uasis norm open.
The following theorem states that almost uniformly asymptotically stable nonhomogeneous chains of Markov operators are generic. Its proof may be par- tially derived from Theorem 3, but for the convenience of the reader we give it in full detail.
Theorem 4. S
a.uasis a dense G
δsubset of S in both sup norm and P norm topologies (i.e., in d
n. supand d
n.Prespectively).
Proof. First we will show that S
a.uasis a d
n. supdense subset of S (the denseness in the metric d
n.Pfollows from the fact that d
n. supis stronger than d
n.P). To this end, given an arbitrary P ∈ S and 0 < ε < 1, consider a convex combination
P
εm,m+1= (1 − ε) P
m,m+1+ εE,
where as before E = (E
m,m+1)
m≥0∈ S is such that E
m,m+1= E for every m ∈ N
0and Ef = ( R
X
f dµ)g for some fixed g ∈ D and any f ∈ L
1(µ). Clearly, P
ε∈ S and d
n. sup(P, P
ε) < 2ε. To prove that P
εis almost uniformly asymptotically stable notice that for any densities f and g ∈ D we have
P
εn−1,nf − P
εn−1,ng
1
= (1 − ε)
P
n−1,nf − P
n−1,ng
1= (1 − ε)
P
n−1,n(f − g)
1≤ (1 − ε) kf − gk
1and therefore
kP
εm,nf − P
εm,ngk
1=
P
n−1,nP
m,n−1f − P
m,n−1g
1≤ (1 − ε)
P
m,n−1f − P
m,n−1g
1. Iterating the last inequality for any f, g ∈ D we have
kP
εm,nf − P
εm,ngk
1≤ (1 − ε)
n−mkf − gk
1. Hence
kP
εm,nf − P
εm,ngk
1≤ 2(1 − ε)
n−mfor any f, g ∈ D. Thus,
sup
f,g∈D
kP
εm,nf − P
εm,ngk
1≤ 2 (1 − ε)
n−m.
Therefore,
n→∞
lim sup
f,g∈D
kP
εm,nf − P
εm,ngk
1= 0
and the denseness of the set S
a.uasin S is proved for both sup norm and P norm topologies.
To show G
δ-ness of S
a.uasobserve that P
m,n+1f − P
m,n+1g
1
=
P
n,n+1(P
m,nf ) − P
n,n+1(P
m,ng)
1≤ kP
m,nf − P
m,ngk
1,
which means that the sequence kP
m,nf − P
m,ngk
1is nonincreasing. It follows that for the fixed m the sequence sup
f,g∈DkP
m,nf − P
m,ngk
1is nonincreasing as well. We obtain that
S
a.uas= (
P ∈ S : ∀
m∈N0lim
n→∞
sup
f,g∈D
kP
m,nf − P
m,ngk
1= 0 )
=
∞
\
m=0
∞
\
k=1
∞
[
n=m+1
(
P ∈ S : sup
f,g∈D
kP
m,nf − P
m,ngk
1< 1 k
)
.
Note that for fixed m < n the function S 3 P 7→ sup
f,g∈D
kP
m,nf − P
m,ngk
1is d
n.Pcontinuous. Hence S
a.uasis a G
δset for the metric d
n.P(and it is G
δset for the stronger metric d
n. sup).
The following example shows that (in contrast to the homogeneous case) in the class of nonhomogeneous chains of Markov operators the set S
a.uasis not open (even though it is a dense G
δsubset of S ).
Example 5. As before, since the measure µ is σ-finite, there exists a sequence {B
k}, B
k∈ A, such that B
i∩B
j= ∅ for i 6= j and X = S
∞k=0
B
k, 0 < µ (B
k) < ∞ for all k ∈ N
0. Define
S
m,m+1f =
m+11Z
X
f dµ
g
m+1+
1 −
m+11X
∞j=m+1
a
j−m−1(f ) · g
j+1for any f ∈ L
1(µ) and any g
j∈ D such that supp g
j⊆ B
jand where a
j(f ) = R
Bj
f dµ. Applying Theorem 3 we obtain that S ∈ S
a.uasas for every f , h ∈ D we have kS
m,m+1f ∧ S
m,m+1hk
1≥
m+11and P
∞m=0 1
m+1
= ∞. Fix ε > 0 and choose M such that
M +12< ε. Consider T ∈ S defined for f ∈ L
1(µ) as follows:
T
m,m+1f =
( S
m,m+1f, if m < M,
P
∞j=m+1
a
j−m−1(f ) · g
j+1, if m ≥ M.
Notice that for m ≥ M we have sup
f,h∈DkT
m,nf − T
m,nhk
1= 2 (e.g. take f , h ∈ D such that supp f ⊆ B
M +1and supp h ⊆ B
M +2). Hence T / ∈ S
a.uas. We easily find that d
n.sup(S, T) ≤
m+12< ε. It follows that S
a.uasis not norm open.
3. Strong operator topology asymptotic stability
This section is dedicated to the study of the asymptotic stability of nonhomo- geneous chains of Markov operators in strong operator topology on S . Similar to the previous section, we introduce two types of limit behaviour with the only difference that the mode of convergence is strong. We begin with
Definition. A nonhomogeneous chain of Markov operators P is called strong
asymptotically stable if there exists (a unique) f
∗∈ D such that for every m ∈ N
0and every f ∈ D
n→∞
lim kP
m,nf − f
∗k
1= 0.
The set of all strong asymptotically stable Markov operators is denoted by S
sas. Obviously S
uas⊆ S
sas.
Theorem 6. The set S
sascof all Markov operators which are not strong asymp- totically stable is P sup topology dense subset of S (i.e., in d
so. sup). Moreover, S
sasccontains P P dense G
δset (i.e., in d
so.P).
Proof. Similar arguments to those which were used towards the proof of The- orem 1 imply the first part of the above theorem. Therefore we only prove the second statement.
Let {f
0, f
1, . . .} be a fixed countable and linearly dense subset of D. As before, we find a sequence {B
k} such that B
k∈ A, B
i∩ B
j= ∅ for i 6= j and X = S
∞k=0
B
k, 0 < µ(B
k) < ∞ for all k ∈ N
0. To see that S
sasccontains the P P dense G
δset observe that
S
sasc⊇ (
P ∈S :∀
t∈N0∀
m∈N0∀
j∈N0∀
ı∈N∀
N ∈N0∃
n>max{N,m}t
X
k=0
Z
Bk
P
m,nf
jdµ < 1 ı )
=
∞
\
t=0
∞
\
m=0
∞
\
j=0
∞
\
ı=1
∞
\
N =0
[
n>max{N,m}
(
P ∈ S :
t
X
k=0
Z
Bk
P
m,nf
jdµ < 1 ı
) .
Clearly, the mapping
S 3 P 7→
t
X
k=0
Z
Bk
P
m,nf
jdµ
is d
so.Pcontinuous, hence the sets {P ∈ S : P
tk=0R
Bk
P
m,nf
jdµ <
1l} are open in the topology induced by d
so.P. It remains to show that the set
(
P ∈ S : ∀
t∈N0∀
m∈N0∀
j∈N0∀
ı∈N∀
N ∈N0∃
n>max{N,m}t
X
k=0
Z
Bk
P
m,nf
jdµ < 1 ı
) (∗)
is P P dense. Now then, let P ∈ S and 0 < ε < 1 be taken arbitrarily. There exists m
0∈ N
0such that
∞
X
l=0
∞
X
m=m0+1
1 2
m+l+1=
∞
X
l=0
1 2
l+1∞
X
m=m0+1
1 2
m= 1
2
m0< ε
2 .
Let g
k∈ D be such that supp g
k= {x ∈ X : g
k(x) 6= 0} ⊆ B
kfor any k ∈ N
0. Consider P
ε∈ S defined as follows:
P
εm,m+1f =
( (1 − ε) P
m,m+1f + ε R
X
f dµ g
m0+1, if 0 ≤ m ≤ m
0, R
X
f dµ g
m+1, if m > m
0for any f ∈ L
1(µ). We notice that Z
t
S
k=0
Bk
P
εm,nf dµ = 0
if n > max{m
0+ 1, t}, so clearly P
ε= (P
εm,m+1)
m≥0is an element of the consid- ered subset (∗). Moreover, we have
d
so.P(P
ε, P)
=
∞
X
m,l=0
1 2
m+l+1P
εm,m+1f
l− P
m,m+1f
l1
=
∞
X
l=0 m0
X
m=0
1 2
m+l+1P
εm,m+1f
l− P
m,m+1f
l1
+
∞
X
l=0
∞
X
m=m0+1
1 2
m+l+1P
εm,m+1f
l− P
m,m+1f
l1
=
∞
X
l=0 m0
X
m=0
1 2
m+l+1(1 − ε)P
m,m+1f
l+ ε
Z
X
f
ldµ
g
m0+1− P
m,m+1f
l1
+
∞
X
l=0
∞
X
m=m0+1
1 2
m+l+1Z
X
f
ldµ
g
m+1− P
m,m+1f
l1
= ε
∞
X
l=0 m0
X
m=0
1 2
m+l+1g
m0+1− P
m,m+1f
l1
+
∞
X
l=0
∞
X
m=m0+1
1 2
m+l+1g
m+1− P
m,m+1f
l1
≤ ε
∞
X
l=0 m0
X
m=0
1
2
m+l+1· 2 +
∞
X
l=0
∞
X
m=m0+1
1 2
m+l+1· 2
< ε
∞
X
l=0 m0
X
m=0
1 2
m+l+ ε
2 · 2 = ε
1 − 1
2
m0 ∞X
l=0
1
2
l+ ε < 2ε + ε = 3ε,
which completes the proof.
On the other hand, the fact that S
uas⊆ S
sasand the Proposition 2 lead to the following
Proposition 7. The set S
sasis P P topology dense in S (i.e., in d
so.P).
Let us proceed with
Definition. A nonhomogeneous chain of Markov operators P is called strong almost asymptotically stable if for every m ∈ N
0and f , g ∈ D
n→∞
lim kP
m,nf − P
m,ngk
1= 0.
The set of all strong almost asymptotically stable Markov operators is denoted by S
a.sas.
Clearly, S
a.uas⊂ S
a.sas. It should be emphasized that unlike the uniform case, in the class of homogeneous chains of Markov operators notions of strong asymptotic stability and strong almost asymptotic stability are essentially different (cf. [4]).
We easily obtain that strong almost asymptotically stable nonhomogeneous chains of Markov operators are generic.
Theorem 8. The set S
a.sasis a dense G
δsubset of S in both P sup and P P strong operator topologies (i.e., in d
so. supand d
so.Prespectively).
Proof. It remains to show the G
δ-ness of S
a.sas. Let {f
0, f
1, . . .} be a fixed countable and linearly dense subset of D. Notice that
S
s.aas= n
P ∈ S : ∀
m∈N0∀
i∈N0∀
j∈N0lim
n→∞
kP
m,nf
i− P
m,nf
jk
1= 0 o
=
∞
\
m=0
∞
\
i=0
∞
\
j=0
∞
\
ı=1
∞
\
N =0
[
n>max{N,m}